Method and system for constructing anti-cancer drug sensitivity prediction model
Through the method based on collaborative filtering recommendation algorithm and matrix decomposition, the problem of insufficient number of data points and accuracy of anti-cancer drug sensitivity prediction models in the prior art is solved, and more efficient and accurate drug response prediction is achieved, providing good results biogenomic interpretability and model interpretability.
Patent Information
- Application Number
- CN202411193942.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-08-28
AI Technical Summary
Existing machine learning models have problems such as limiting the number of data points and insufficient accuracy of drug response prediction when predicting the sensitivity of anti-cancer drugs. It is difficult to effectively identify biomarkers of drug responses, and the accuracy and sensitivity of detection technology are insufficient.
Using a method based on a collaborative filtering recommendation algorithm, the drug reaction matrix is constructed and decomposed into deviations, cell line matrix and drug matrix. The drug and cell line data are projected into high-dimensional latent space using matrix decomposition. Combined with Bayesian sigmoid curves to fit IC50 values, the sum of squared errors is optimized to train the drug response calculation model, and the model performance is evaluated using five-fold cross-validation and Pearson correlation.
Improves the efficiency and accuracy of anti-cancer drug sensitivity prediction, provides good biogenomic interpretability, can perform well on untouched data sets, and can determine the intensity and directionality of drug-to-pathway association, improves The accuracy of drug response prediction and the interpretability of the model are carried out.
Smart Images

Figure CN119069144A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of deep learning research and the technical field of developing computational models using computers, and in particular to a method for improving the performance of a machine learning model based on a machine learning algorithm and thereby predicting the sensitivity of anticancer drugs. Background Art
[0002] Cancer is a highly complex and diverse disease in which tissue and organ dysfunction occurs due to abnormal cell proliferation and malignant transformation. Normally, normal cells proliferate and divide to a limited extent according to the needs of the body, and die in an orderly manner after completing their functions. However, cancer cells proliferate and spread without limit, forming malignant tumors. Cancer can occur in any part of the body's tissues and organs, seriously affecting the patient's quality of life and potentially endangering their lives. The era of precision medicine pursues personalized medicine, and the basic goal is to use targeted drugs to design and determine personalized therapies to maximize drug efficacy while minimizing adverse side effects. How to accurately calculate and predict the response of cancer patients to various targeted drug therapies based on the patient's molecular and clinical characteristics has become a crucial basic issue in the era of precision medicine.
[0003] In recent years, the number of patients with molecular data and tumor bioinformation in cancer information has been increasing. At the same time, large-scale studies such as The Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium (ICGC) have integrated data information, and cancer-driven gene identification research has benefited greatly. However, these data sources only contain genetic information of cancer groups, and usually lack corresponding drug response information, so they are not suitable for identifying biomarkers of drug response. On the other hand, many research works, including the CCLE Cancer Cell Line Encyclopedia and the Genomics of Drug Sensitivity in Cancer Collaborative (GDSC) projects, have screened drugs on multiple groups of cancer cell lines. These cell line datasets enable research to use genomic information features, apply mathematical and statistical methods to decipher the functional relationship between drug response and genomic information, and build models that can predict accurate patient-specific drug responses.
[0004] Researchers have proposed various types of models that make full use of genomic features to predict drug response. Among them, the most widely used model is the drug-specific model. This model is based on the genetic information of the cell line and the drug response information of each drug tested separately, and each drug is trained independently. Models belonging to this category of methods generally use data sets with similar structures, most of which are based on baseline gene expression or a combination of gene expression and other genomic information (such as copy number changes and DNA methylation). There are many ways to structure the model, including linear regression models, neural networks, random forests, support vector machines and kernel regression based on multiple genomic information, and neural network models that include drug attribute information. Drug-specific models are usually limited by the number of cell lines tested with a given drug. In order to increase the number of data points and obtain a more robust and general drug sensitivity model, Costello et al. proposed the Bayesian Multi-Task Multi-Kernel Learning (BMTMKL) method and showed the best performance in the DREAM challenge for drug response prediction. This work emphasizes the importance of sharing information across drugs to improve the accuracy of drug response prediction.
[0005] Although the above studies have used many types of methods to successfully predict the accuracy of anticancer drugs, the latest version of the "Guidelines for the Clinical Application of New Anti-tumor Drugs (2020 Edition)" proposes that the use of anti-tumor drugs should not only be based on the results of pathological diagnosis, but also follow the principle that targeted drug therapy can only be used after genetic testing, that is, targeted therapy, testing first. Precision diagnosis and treatment of cancer is inseparable from the development of detection technology and its widespread clinical application. How to improve the accuracy and sensitivity of detection technology is crucial for precision diagnosis and treatment. Machine learning has the ability to learn knowledge from data and experience, and provides important technical support for many interdisciplinary disciplines. Bioinformatics research includes data acquisition, data management, data analysis, simulation experiments and other links. Machine learning can be used to extract necessary knowledge from large amounts of biological data for learning and prediction.
[0006] Anticancer drug sensitivity testing is the application of computer technology to the field of biology to improve the prediction of anticancer drug sensitivity and the identification of biomarkers related to drug response, thereby accelerating drug discovery and saving the lives of cancer patients. Based on the current existing research, the performance of machine learning models is improved, thereby improving the efficiency and accuracy of anticancer drug sensitivity prediction. Summary of the invention
[0007] The purpose of the present invention is to provide a method for predicting drug sensitivity based on the performance of a machine learning model in order to address the technical defects in predicting drug sensitivity in the prior art.
[0008] The technical solution adopted to achieve the purpose of the present invention is:
[0009] A method for constructing an anticancer drug sensitivity prediction model comprises the following steps:
[0010] Step 1: Build and preprocess the dataset:
[0011] Step 1-1, the data set comes from CCLE and GDSC studies;
[0012] Step 1-2, recalculating the IC50 values of the two data sets;
[0013] Step 1-3, comparing the IC50 values obtained from the two data sets to determine the final data set;
[0014] Step 2: Build a model for drug response calculation:
[0015] Step 2-1, based on the idea of using latent factors for feature extraction in the collaborative filtering recommendation algorithm, the drug response matrix W is constructed and decomposed into the deviation S, the cell line matrix C and the drug matrix M;
[0016] Step 2-2, use the decomposed drug matrix M to learn the drug genome space, project the drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the cell line-specific drug response;
[0017] Step 2-3, construct the relationship between the drug response matrix W, the cell line matrix C and the drug matrix M to determine the potential correlation between the cell line and the drug genome;
[0018] Step 3, training and evaluating the drug response computational model;
[0019] The optimized sum of squared error loss function is selected as the optimization algorithm for the drug response calculation model; wherein the sum of squared error loss function is:
[0020]
[0021]
[0022] in and W ij are the observed and predicted sensitivity scores of cell line i using drug j, respectively, and the drug response matrix is the number of drug response experiments in the training dataset;
[0023] The method for optimizing the sum of square error loss function is as follows: optimizing the sum of square error loss function by using gradient descent; obtaining all parameters in the drug response matrix W according to the measured drug response values; and M P , MQ The gradient function is derived as follows:
[0024]
[0025] In the drug response computational model, initialization is 0, M P and M Q Using the minimum random number, perform batch gradient descent using the above gradient calculation formula to optimize M P and M Q .
[0026] Then, the evaluation method of the drug response calculation model was determined, specifically: using five-fold cross validation, the drug response calculation model results were compared with the correlation of the known response of each drug; Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained; at the same time, the normalized discounted cumulative gain (NDCG) was used to evaluate the drug response calculation model of each cell line, and the calculation formula is as follows:
[0027]
[0028] where g is the predicted rank of the drug tested on the cell line, w is a list of observed drug sensitivity scores, It is the ranking of known drugs calculated based on measured drug response values; NDCG values range from 0 to 1, where 1 means that the drug response calculation model correctly predicts the ranking of the drug.
[0029] Furthermore, in step 1-1, the GDSC and CCLE datasets are used as benchmark datasets.
[0030] Furthermore, in step 1-2, for the two data sets of CCLE and GDSC, the IC50 value was selected as the indicator of drug sensitivity, and the IC50 value was recalculated using the Bayesian sigmoid curve fitting method;
[0031] In CCLE, each cell line was tested with 6 different drug concentrations, and a cell activity value AV was measured at each concentration; in GDSC, 7 different drug concentrations were used to measure a cell activity value AV;
[0032]
[0033] Wherein, T, P, and N represent the intensity values measured in the test drug, positive control, and negative control, respectively;
[0034] The JAGS method is used to fit the Bayesian sigmoid curve to obtain the dose-response fitting curve for each pair of drugs and cell lines:
[0035]
[0036] Define a as the cell line ID, b as the drug ID, x ab is the logarithmic scale dose, y ab As a result of the response to a given drug and cell line pair, represents precision, which is the inverse of the variance, Represents the partial derivative of the input variable; using parameter γ 0b and γ 1b Two parameters are used to simulate the control scale parameter θ distribution; where γ 0b Represents the center of the drug response curve, i.e., the IC50 value, γ 1b Used to capture the steepness of the curve.
[0037] Furthermore, in steps 1-3, the final data set is determined by recalculating the IC50 value; after recalculating the IC50 value, the paired differences of the measured variables are first measured using a paired p test, and assuming that the difference between the two IC50 values is 0, the obtained p value is less than 0.01, and a decision can be made not to reject the null hypothesis.
[0038] Furthermore, in step 2-1, the relationship between cell lines and the dependency between drugs are analyzed using a collaborative filtering algorithm to identify new drug-response associations. Suppose a user {u1,u2,...u m} and n items {h1,h2,...h n}, each user u1 has a corresponding item list Used to record the user's rating data for these items, or record the user's preference data inferred by the model through the user's behavior.
[0039] Furthermore, in step 2-2, matrix decomposition is used to learn the drug genome space, and the drug and cell line data are projected into a high-dimensional latent space. The dot product between the cell line vector and the drug vector is used to provide the specific drug response of the cell line, and the drug response model is calculated according to the following formula:
[0040]
[0041] M P is a transformation matrix that transforms the cell feature x i Projected into the latent space, are the deviation terms of drug j and cell line i, respectively, and k j and d iis the vector of drug j and cell line i in the n-dimensional latent space, M P is a transformation matrix that transforms the cell feature x i Projection into latent space.
[0042] Furthermore, in order to better train the model in step 3, a drug sensitivity scoring coefficient w=-log(IC50) is defined. The higher the score, the more sensitive the cell line is to the drug. Through performance testing, the model can extract more biological features to predict the reaction results, making the results more accurate.
[0043] The present invention also provides a system for constructing an anticancer drug sensitivity prediction model, comprising:
[0044] The dataset construction module is used to construct the dataset and preprocess it, including the following processes:
[0045] Step 1-1, the data set comes from CCLE and GDSC studies;
[0046] Step 1-2, recalculating the IC50 values of the two data sets;
[0047] Step 1-3, comparing the IC50 values obtained from the two data sets to determine the final data set;
[0048] The model building module is used to build a model for drug response calculation, including the following processes:
[0049] Step 2-1, based on the idea of using latent factors for feature extraction in the collaborative filtering recommendation algorithm, the drug response matrix W is constructed and decomposed into the deviation S, the cell line matrix C and the drug matrix M;
[0050] Step 2-2, use the decomposed drug matrix M to learn the drug genome space, project the drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the cell line-specific drug response;
[0051] Step 2-3, construct the relationship between the drug response matrix W, the cell line matrix C and the drug matrix M to determine the potential correlation between the cell line and the drug genome;
[0052] Training module, used to train and evaluate drug response computational models;
[0053] The optimized sum of squared error loss function is selected as the optimization algorithm for the drug response calculation model; wherein the sum of squared error loss function is:
[0054]
[0055]
[0056] in and W ij are the observed and predicted sensitivity scores of cell line i using drug j, respectively, and the drug response matrix is the number of drug response experiments in the training dataset;
[0057] The method for optimizing the sum of square error loss function is as follows: optimizing the sum of square error loss function by using gradient descent; obtaining all parameters in the drug response matrix W according to the measured drug response values; and M P , M Q The gradient function is derived as follows:
[0058]
[0059] In the drug response computational model, initialization is 0, M P and M Q Using the minimum random number, perform batch gradient descent using the above gradient calculation formula to optimize M P and M Q .
[0060] Then, the evaluation method of the drug response calculation model was determined, specifically: using five-fold cross validation, the drug response calculation model results were compared with the correlation of the known response of each drug; Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained; at the same time, the normalized discounted cumulative gain (NDCG) was used to evaluate the drug response calculation model of each cell line, and the calculation formula is as follows:
[0061]
[0062] where g is the predicted rank of the drug tested on the cell line, w is a list of observed drug sensitivity scores, It is the ranking of known drugs calculated based on measured drug response values; NDCG values range from 0 to 1, where 1 means that the drug response calculation model correctly predicts the ranking of the drug.
[0063] Furthermore, in step 1-1, the GDSC and CCLE datasets are used as benchmark datasets.
[0064] Furthermore, in step 1-2, for the two data sets of CCLE and GDSC, the IC50 value was selected as the indicator of drug sensitivity, and the IC50 value was recalculated using the Bayesian sigmoid curve fitting method;
[0065] In CCLE, each cell line was tested with 6 different drug concentrations, and a cell activity value AV was measured at each concentration; in GDSC, 7 different drug concentrations were used to measure a cell activity value AV;
[0066]
[0067] Wherein, T, P, and N represent the intensity values measured in the test drug, positive control, and negative control, respectively;
[0068] The JAGS method is used to fit the Bayesian sigmoid curve to obtain the dose-response fitting curve for each pair of drugs and cell lines:
[0069]
[0070] Define a as the cell line ID, b as the drug ID, x ab is the logarithmic scale dose, y ab As a result of the response to a given drug and cell line pair, represents precision, which is the inverse of the variance, Represents the partial derivative of the input variable; using γ 0b and γ 1b Two parameters are used to simulate the control scale parameter θ distribution; where γ 0b Represents the center of the drug response curve, i.e., the IC50 value, γ 1b Used to capture the steepness of the curve.
[0071] Compared with the prior art, the present invention has the following beneficial effects:
[0072] 1. The present invention not only uses the predicted correlation of drug response for each cell line as an evaluation indicator, but also uses the weighted ranking indicator (NDCG) often used in recommendation systems to evaluate the model's ability to correctly rank drugs for a certain cell line. Under these two evaluation indicators, the recommendation system-based model can always provide good results and perform well on data sets that the model has never been exposed to.
[0073] 2. The recommendation system-based model of the present invention has very good biogenomic interpretability. The trained model projects cell lines and drugs into the drug genome space, establishes feature vectors of drugs and cell lines in the latent space, and can be used to explore drug-drug, cell line-cell line and drug-cell line relationships, conduct visualization research and genomic law exploration and biological interpretation. The recommendation system-based model enables this embodiment to determine the strength and direction of the association between drugs and pathways, and provide guidance for future research directions. BRIEF DESCRIPTION OF THE DRAWINGS
[0074] Figure 1 is a flow chart of the present invention;
[0075] Figure 2 Schematic diagram of the correlation between the IC50 values recalculated in the embodiment and different data sets;
[0076] Figure 3 This is an example diagram of the Spearman correlation between CCLE and GDSC data sets of three drugs in the embodiment;
[0077] Figure 4 Schematic diagram of the lack of specificity of drugs with a median IC50 < 1 μm in the embodiment; The lack of specificity of drugs with a median IC50 < 1 μm, (A) CCLE data set (B) GDSC data set;
[0078] Figure 5 This is an example diagram of the drug genome potential space in the embodiment;
[0079] Figure 6 Similar performance result graphs were generated for the 10 models trained based on different random starting points in the embodiment;
[0080] Figure 7 Schematic diagram of the dimension of the latent space and the model accuracy in the embodiment; DETAILED DESCRIPTION
[0081] The present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0082] This embodiment provides a method for constructing an anticancer drug sensitivity prediction model, comprising the following steps:
[0083] Step 1: Build and preprocess the dataset
[0084] Step 1-1: Prepare a data set. The data set comes from two large-scale studies: CCLE (Cancer Cell Line Encyclopedia) and GDSC (Genomics of Drug Sensitivity in Cancer (GDSC)).
[0085] In step 1-1, the CCLE project described is a research project aimed at collecting and sharing genomic information and drug sensitivity data of various cancer cell lines. A large-scale genomic data set was generated for 947 human cancer cell lines, and 24 compounds were pharmacologically analyzed in about 500 of these cell lines, generating pharmacological maps of 24 compounds (targeted and cytotoxic agents) across 479 cell lines, accelerating the progress of cancer biology and drug discovery.
[0086] In step 1-1, the GDSC project described is a project aimed at studying the relationship between cancer drug sensitivity and genomics. It is a larger dataset developed by Wanjuan Yang et al. in 2012, containing drug sensitivity data from nearly 75,000 experiments, describing the response of nearly 700 cancer cell lines to 138 anticancer drugs. In order to determine the molecular markers of drug response, the GDSC dataset was combined with a large genomic dataset obtained from the COSMIC (the Catalogue of Somatic Mutations in Cancer) database, which summarizes somatic mutation information of cancer genes, gene amplification and deletion, tissue type and transcription data, etc.
[0087] Both CCLE and GDSC datasets are rich in genomic information data, and there is considerable concordance between the two in terms of drug response measures and predictive features of these responses.
[0088] Step 1-2 recalculates the IC50 value of the data set in step 1-1;
[0089] In step 1-2, the IC50 (half maximal inhibitory concentration) value was selected as an indicator of drug sensitivity. IC50 is expressed as a concentration, and the higher the drug concentration required to kill the cell line, the higher the drug resistance of the cell line. If the IC50 is very low, then the drug concentration (or amount) required to kill the cell line is also very low, which means that the cell line is more sensitive. For both CCLE and GDSC datasets, the study retained all cell lines with baseline gene expression data. In order to make the values between datasets more comparable, the IC50 values were recalculated using the Bayesian sigmoid curve fitting method.
[0090] (1) The dose-response experimental data of CCLE and GDSC were observed. In CCLE, each cell line was tested with 6 different drug concentrations, and a cell activity value (AV) was measured at each concentration. In GDSC, 7 different drug concentrations were used to measure a cell activity value (AV).
[0091]
[0092] Wherein, T, P and N represent the intensity values (cell counts) measured in the test drug, positive control and negative control, respectively.
[0093] (2) Use the JAGS (Just another Gibbs sampler) method to fit the Bayesian sigmoid curve and obtain the dose-response fitting curve for each pair of drug and cell line. In simple terms, define a as the cell line ID, b as the drug ID, and x as ab is the logarithmic scale dose, y ab is the response result for a given drug and cell line pair. Using the parameter γ 0b and γ 1b Two parameters to simulate the Student's distribution. Among them, γ 0b represents the center of the drug response curve (i.e., IC50 value), γ 1b It is used to capture the steepness of the curve. The fitting formula obtained is as follows:
[0094]
[0095] In addition, assuming that the drug response curve γ 0b and γ 1b The two parameters of obey Normal and Gamma distribution respectively.
[0096]
[0097] For the remaining parameters, the theory described by Kruschke was used, using non-informative priors. After the fitting calculations were completed, the IC50 values of the cells were recalculated using the fitted curves.
[0098] Step 1-3 compares the IC50 values obtained in step 1-2 to determine the final data set: After recalculating the IC50 values, in order to better carry out subsequent work, the study compared the two IC50 values. First, the paired p test was used to measure the paired differences of the variables. Assuming that the difference between the two IC50 values is 0, the p value obtained is less than 0.01, and a decision can be made not to reject the null hypothesis. In addition, the study then calculated the correlation of the IC50 values of the two data sets. Figure 2 It can be seen that the Spearman correlations between each drug are higher than the IC50 values provided when using the IC50 values recalculated from the original dose-response data. Figure 3 An example of Spearman correlation between CCLE and GDSC datasets for three drugs is shown, with the processed IC50 values resulting in increased correlation between the data in the two datasets. As described above, the recalculated IC50 values are more comparable between the different datasets.
[0099] All cell lines showed very high drug sensitivity to drugs with low median IC50, which reflects that these drugs may lack specificity (low rank entropy) and be cytotoxic. Figure 4As shown, through data statistics in the two data sets (A) CCLE and (B) GDSC, it is found that drugs with a median IC50 < 1 μM are often cytotoxic drugs and have consistently high toxicity in all cell lines. Their presence makes the drug response prediction problem easier, so this example excludes them and focuses on predicting the response of targeted anticancer drugs.
[0100] In the examples of the present invention, after screening, 492 cell lines of CCLE and 983 cell lines of GDSC were obtained.
[0101] Step 2: Construct a model for drug response calculation (the model for drug response calculation uses a collaborative filtering recommendation algorithm, hereinafter referred to as the model)
[0102] The main idea of the model is that the drug response matrix W is decomposed into the bias matrix S, the cell line matrix C and the drug matrix M.
[0103] Step 2-1 constructs a drug response matrix and decomposes it into a cell line matrix and a drug matrix. This embodiment conducts experiments based on the idea of using latent factors for feature extraction in a collaborative filtering recommendation algorithm.
[0104] Recommendation algorithms can be simply summarized as algorithms that recommend the most suitable items, products or services to specific users. They include collaborative filtering, content-based, knowledge-based and hybrid methods. As one of the most successful methods for building recommendation systems, collaborative filtering (CF) technology is a widely used technology in recommendation systems that uses the known preferences of a group of users to recommend or predict the unknown preferences of other users.
[0105] So far, collaborative filtering algorithms can be roughly divided into three types: ① memory-based collaborative filtering algorithms, among which the neighborhood-based collaborative filtering algorithm is the most popular, ② model-based collaborative filtering algorithms, and ③ collaborative filtering technology based on the matrix decomposition model of cell line and drug similarity.
[0106] This embodiment combines the above three collaborative filtering algorithms to analyze the relationship between cell lines and the dependency between drugs to identify new drug response associations.
[0107] Step 2-2 builds a model for calculating drug response: Use matrix decomposition to learn the "drug genome space" (drug genome space refers to a high-dimensional space in which each dimension represents a biological feature related to drug response, such as drug properties (such as chemical structure, drug mechanism, etc.) and genetic and phenotypic characteristics of cell lines (such as gene expression, mutation status, etc.)), project drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the specific drug response of the cell line. The drug response model is calculated according to the following formula:
[0108]
[0109] k j =y j M Q (6)
[0110] d i =x i M P (7)
[0111] Substituting (5), (6) and (7) into the transformation, we get the simplified result as shown in formula (8):
[0112]
[0113] W ij is the model-predicted sensitivity score of cell line i to drug j, μ is the overall average drug response, are the deviation terms of drug j and cell line i, respectively, and k j and d i is the vector of drug j and cell line i in the n-dimensional latent space. Q is a matrix containing drug vectors, and y is a drug indicator vector. P is a transformation matrix that transforms the cell feature x i Projected to latent space. Based on the cross-validation performance, the value of n is set to 12 for CCLE and GDSC datasets.
[0114] Step 2-3 simplifies formula (8) to obtain the following formula:
[0115] W=S+CM T (9)
[0116] The cell line matrix C can be viewed as a vector of cell line vectors in the latent space, and the drug matrix M can be viewed as a vector of drug vectors in the latent space. The latent drug genome space is able to capture the interaction between the drug and the genomic background of the cell line, so the dot product (k, d) between the cell line vector and the drug vector can represent the interaction between the drug and the cell line, as shown in the attached figure. Figure 5As shown, using vectors to represent the characteristic information of cell lines and drugs, it is easy to see that cell line i1 is sensitive to drugs j1 and j2, but not to drug j3. Similarly, cell line i2 is different from cell line i1, and it does not respond to drugs j1 and j2. This representation can predict the drug response of unseen cell lines or patient samples.
[0117] Step 3: Train and evaluate the drug response computational model
[0118] Step 3-1: Training a model for drug response calculation
[0119] This embodiment is optimized using the "sum of squared errors" loss function described below.
[0120]
[0121] in and W ij are the observed and predicted sensitivity scores of cell line i using drug j, is the number of drug response experiments in the training dataset.
[0122] Step 3-2 Optimize the loss function
[0123] All parameters in W are obtained based on the measured drug response values by optimizing the loss function in step 3-2 using gradient descent. M P , M Q The gradient function is derived as follows:
[0124]
[0125]
[0126] In model training, initialization is 0, M P and M Q Using the minimum random number, perform batch gradient descent using the above gradient calculation formula to optimize M P and M Q .
[0127] Step 3-3 Determine the model evaluation method
[0128] A five-fold cross validation was performed on the data set in step 2-2, and the correlation (or squared difference) of the model results with the known response of each drug (across cell lines) was compared. The Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained. In addition, this example also uses the normalized discounted cumulative gain (NDCG Normalized Discounted Cumulative Gain) and the discounted cumulative gain (DCG Discounted Cumulative Gain) to evaluate the model of each cell line. The calculation formula is as follows:
[0129]
[0130] where g is the predicted rank of the drug tested on the cell line, w is a list of observed drug sensitivity scores, It is the rank of known drugs calculated based on the measured drug response values. NDCG takes values ranging from 0 to 1, where 1 means that the model correctly predicted the rank of the drug. The numerator in DCG is designed to give more weight to drugs with higher sensitivity scores, while the denominator gives priority to drugs predicted to have a higher rank.
[0131] Step 3-4 Predict the performance of the model
[0132] In order to better train the model, the study defined a drug sensitivity scoring coefficient w = -log (IC50). The higher the score, the more sensitive the cell line is to the drug.
[0133] In this example, the model is trained on the complete dataset to obtain the unique features of drugs and cell lines, and then the drug genome potential space of drug-drug, cell line-cell line and drug-cell line associations in CCLE and GDSC datasets is derived.
[0134] In order to test the robustness of the model, this example uses the same data set and builds 10 different models from different random starting points (S). The gradient descent method is used for optimization. The models are evaluated using normalized discounted cumulative returns. The results are shown in the box plot as shown in the attached figure. Figure 6 As shown. It can be seen that the average NDCG of these models is stable at around 0.94, showing very similar performance, proving the repeatability of the model. In order to obtain better training results, the dimension of the latent space was changed, and multiple groups of comparative experiments were conducted using five-fold cross-validation to explore the impact of the number of dimensions on the accuracy of the model. The Pearson correlation coefficient between the IC50 value obtained by the model and the true value was calculated, and the results are represented by scatter points as shown in the attached figure. Figure 7As shown in the figure, as the dimension of the latent space increases, the correlation between the predicted results and the true values continues to increase, and the model can extract more biological features to predict the reaction results, making the results more accurate.
Claims
1. A method for constructing a prediction model for anticancer drug sensitivity. It is characterized in that The following steps are included: Step 1: Build and preprocess the dataset: Step 1-1, the data set comes from CCLE and GDSC studies; Step 1-2, recalculating the IC50 values of the two data sets; Step 1-3, comparing the IC50 values obtained from the two data sets to determine the final data set; Step 2: Build a model for drug response calculation: Step 2-1, based on the idea of using latent factors for feature extraction in the collaborative filtering recommendation algorithm, the drug response matrix W is constructed and decomposed into the deviation S, the cell line matrix C and the drug matrix M; Step 2-2, use the decomposed drug matrix M to learn the drug genome space, project the drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the cell line-specific drug response; Step 2-3, construct the relationship between the drug response matrix W, the cell line matrix C and the drug matrix M to determine the potential correlation between the cell line and the drug genome; Step 3, training and evaluating the drug response computational model; The optimized sum of squared error loss function is selected as the optimization algorithm of the drug response calculation model; wherein the sum of squared error loss function is: in and W ij are the observed and predicted sensitivity scores of cell line i using drug j, respectively, and the drug response matrix is the number of drug response experiments in the training dataset; The method for optimizing the sum of square error loss function is as follows: optimizing the sum of square error loss function by using gradient descent; obtaining all parameters in the drug response matrix W according to the measured drug response values; and M P , M Q The gradient function is derived as follows: in, and are the deviation terms of drug j and cell line i, respectively, and k j and d i is the vector of drug j and cell line i in the n-dimensional latent space; M Q is a matrix containing drug vectors, and y is a drug indicator vector. P is a transformation matrix that transforms the cell feature x i Projection into latent space; in drug response computational models, initialization and is 0, M P and M Q Using the minimum random number, perform batch gradient descent using the above gradient calculation formula to optimize M P and M Q ; Then, the evaluation method of the drug response calculation model was determined, specifically: using five-fold cross validation, the drug response calculation model results were compared with the correlation of the known response of each drug; Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained; at the same time, the normalized discounted cumulative gain (NDCG) was used to evaluate the drug response calculation model of each cell line, and the calculation formula is as follows: where g is the predicted rank of the drug tested on the cell line, w is a list of observed drug sensitivity scores, It is the ranking of known drugs calculated based on measured drug response values; NDCG values range from 0 to 1, where 1 means that the drug response calculation model correctly predicts the ranking of the drug.
2. The method for constructing an anticancer drug sensitivity prediction model according to claim 1, It is characterized in that In step 1-1, the GDSC and CCLE datasets are used as benchmark datasets.
3. The method for constructing an anticancer drug sensitivity prediction model according to claim 2, It is characterized in that In step 1-2, for the two data sets of CCLE and GDSC, the IC50 value was selected as the indicator of drug sensitivity, and the IC50 value was recalculated using the Bayesian sigmoid curve fitting method; In CCLE, each cell line was tested with 6 different drug concentrations, and a cell activity value AV was measured at each concentration; in GDSC, 7 different drug concentrations were used to measure a cell activity value AV; Wherein, T, P, and N represent the intensity values measured in the test drug, positive control, and negative control, respectively; The JAGS method is used to fit the Bayesian sigmoid curve to obtain the dose-response fitting curve for each pair of drugs and cell lines: Define a as the cell line ID, b as the drug ID, x ab is the logarithmic scale dose, y ab As a result of the response to a given drug and cell line pair, represents precision, which is the inverse of the variance, Represents the partial derivative of the input variable; using parameter γ 0b and γ 1b Two parameters are used to simulate the control scale parameter θ distribution; where γ 0b Represents the center of the drug response curve, i.e., the IC50 value, γ 1b Used to capture the steepness of the curve.
4. The method for constructing an anticancer drug sensitivity prediction model according to claim 1, It is characterized in that In steps 1-3, the final data set is determined by recalculating the IC50 value; after recalculating the IC50 value, the paired difference of the measured variables is firstly measured using the paired p test, and assuming that the difference between the two IC50 values is 0, the obtained p value is less than 0.01, and a decision can be made not to reject the null hypothesis.
5. The method for constructing an anticancer drug sensitivity prediction model according to claim 1, It is characterized in that In step 2-1, the relationship between cell lines and the dependency between drugs are analyzed using a collaborative filtering algorithm to identify new drug-response associations. 1 ,u 2 ,……u m } and n items {h 1 ,h 2 ,……h n }, each user u 1 There is a corresponding list of items Used to record the user's rating data for these items, or record the user's preference data inferred by the model through their behavior.
6. The method for constructing an anticancer drug sensitivity prediction model according to claim 1, It is characterized in that In step 2-2, matrix decomposition is used to learn the drug genome space, and the drug and cell line data are projected into a high-dimensional latent space. The dot product between the cell line vector and the drug vector is used to provide the specific drug response of the cell line, and the drug response model is calculated according to the following formula: M P is a transformation matrix that transforms the cell feature x i Projected into the latent space, and are the deviation terms of drug j and cell line i, respectively, and k j and d i is the vector of drug j and cell line i in the n-dimensional latent space, M P is a transformation matrix that transforms the cell feature x i Projection into latent space.
7. The method for constructing an anticancer drug sensitivity prediction model according to claim 1, It is characterized in that In order to better train the model in step 3, the drug sensitivity scoring coefficient w=-log(IC50) is defined. The higher the score, the more sensitive the cell line is to the drug. Through performance testing, the model can extract more biological features to predict the reaction results, making the results more accurate.
8. A system for constructing a prediction model for anticancer drug sensitivity. It is characterized in that include, The dataset construction module is used to construct the dataset and preprocess it, including the following processes: Step 1-1, the data set comes from CCLE and GDSC studies; Step 1-2, recalculating the IC50 values of the two data sets; Step 1-3, comparing the IC50 values obtained from the two data sets to determine the final data set; The model building module is used to build a model for drug response calculation, including the following processes: Step 2-1, based on the idea of using latent factors for feature extraction in the collaborative filtering recommendation algorithm, the drug response matrix W is constructed and decomposed into the deviation S, the cell line matrix C and the drug matrix M; Step 2-2, use the decomposed drug matrix M to learn the drug genome space, project the drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the cell line-specific drug response; Step 2-3, construct the relationship between the drug response matrix W, the cell line matrix C and the drug matrix M to determine the potential correlation between the cell line and the drug genome; Training module, used to train and evaluate drug response computational models; The optimized sum of squared error loss function is selected as the optimization algorithm of the drug response calculation model; wherein the sum of squared error loss function is: in and W ij are the observed and predicted sensitivity scores of cell line i using drug j, respectively, and the drug response matrix is the number of drug response experiments in the training dataset; The method for optimizing the sum of square error loss function is as follows: optimizing the sum of square error loss function by using gradient descent; obtaining all parameters in the drug response matrix W according to the measured drug response values; and M P , M Q The gradient function is derived as follows: In the drug response computational model, initialization and is 0, M P and M Q Using the minimum random number, perform batch gradient descent using the above gradient calculation formula to optimize M P and M Q ; Then, the evaluation method of the drug response calculation model was determined, specifically: using five-fold cross validation, the drug response calculation model results were compared with the correlation of the known response of each drug; Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained; at the same time, the normalized discounted cumulative gain (NDCG) was used to evaluate the drug response calculation model of each cell line, and the calculation formula is as follows: where g is the predicted rank of the drug tested on the cell line, w is a list of observed drug sensitivity scores, It is the ranking of known drugs calculated based on measured drug response values; NDCG values range from 0 to 1, where 1 means that the drug response calculation model correctly predicts the ranking of the drug.
9. The system for constructing an anticancer drug sensitivity prediction model according to claim 8, It is characterized in that In step 1-1, the GDSC and CCLE datasets are used as benchmark datasets.
10. The method for constructing an anticancer drug sensitivity prediction model according to claim 9, It is characterized in that In step 1-2, for the two data sets of CCLE and GDSC, the IC50 value was selected as the indicator of drug sensitivity, and the IC50 value was recalculated using the Bayesian sigmoid curve fitting method; In CCLE, each cell line was tested with 6 different drug concentrations, and a cell activity value AV was measured at each concentration; in GDSC, 7 different drug concentrations were used to measure a cell activity value AV; Wherein, T, P, and N represent the intensity values measured in the test drug, positive control, and negative control, respectively; The JAGS method is used to fit the Bayesian sigmoid curve to obtain the dose-response fitting curve for each pair of drugs and cell lines: Define a as the cell line ID, b as the drug ID, x ab is the logarithmic scale dose, y ab As a result of the response to a given drug and cell line pair, represents precision, which is the inverse of the variance, Represents the partial derivative of the input variable; using γ 0b and γ 1b Two parameters are used to simulate the control scale parameter θ distribution; where γ 0b Represents the center of the drug response curve, i.e., the IC50 value, γ 1b Used to capture the steepness of the curve.
Citation Information
Patent Citations
Genetic material specificity based tumor cell drug sensitivity evaluation method
CN105005693A
Cell line drug response prediction method based on K-nearest neighbor constraint matrix decomposition
CN110739028A
Collaborative anti-tumor multi-drug combination effect prediction method based on deep learning
CN111223577A
Drug sensitivity prediction method based on self-expression model
CN112164474A
Drug sensitivity prediction method based on depth factorization machine
CN112435754A
